AI Paper Atlas

Research map / Large language models / Agents, code and math

PRM-PBE: Process Reward Model for Reinforcement Learning in Programming-by-Example

Yue Fang, Zhi Jin, Jie An, Hongshen Chen, Jiangmeng Li, Xiaohong Chen, Naijun Zhan

ICML 2026 · Large language models › Agents, code and math · approach: Neural program synthesis and transpilation

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add external instruction repair in reasoning tree construction (ablation)PRM-PBE on DeepSeek-Coder-V2 · Lists, MBPP, PROSE averagePass@1 (%): 44.96 -> 49.79Improved
add external instruction repair in reasoning tree construction (ablation)PRM-PBE on Qwen-3 8B · Lists, MBPP, PROSE averagePass@1 (%): 42.55 -> 47.35Improved
add preference learning (pairwise ranking) for the PRM (ablation)PRM-PBE on DeepSeek-Coder-V2 · Lists, MBPP, PROSE averagePass@1 (%): 47.02 -> 49.79Improved
add preference learning (pairwise ranking) for the PRM (ablation)PRM-PBE on Qwen-3 8B · Lists, MBPP, PROSE averagePass@1 (%): 44.55 -> 47.35Improved
add three-stage curriculum learning (ablation)PRM-PBE on DeepSeek-Coder-V2 · Lists, MBPP, PROSE averagePass@1 (%): 47.55 -> 49.79Improved
add three-stage curriculum learning (ablation)PRM-PBE on Qwen-3 8B · Lists, MBPP, PROSE averagePass@1 (%): 44.94 -> 47.35Improved
use process reward instead of the best outcome reward model (ORM-Learned) (ablation)PRM-PBE on DeepSeek-Coder-V2 · five PBE benchmarksPass@1 average (%): 49.29 -> 56.61Improved
use process reward instead of binary outcome reward on D-augmented data (ablation)PRM-PBE on DeepSeek-Coder-V2 · five PBE benchmarksPass@1 average (%): 46.91 -> 56.61Improved
use RL with PRM instead of SFTDeepSeek-Coder-V2 · five PBE benchmarksPass@1 average (%): 42.76 -> 56.61Improved
use RL with PRM instead of SFTLlama-3 · five PBE benchmarksPass@1 average (%): 39.30 -> 54.18Improved
use PRM-PBE instead of GPT-4o WPS promptingDeepSeek-Coder-V2 · five PBE benchmarksPass@1 average (%): 47.49 -> 56.61Improved
use PRM-PBE instead of SFT on MBPP failure casesDeepSeek-Coder-V2 · MBPP failurespartial match errors: 33 -> 20Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy